-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathtrain.py
180 lines (147 loc) · 8.13 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
from operator import imod
import os
import torch
import torch.nn as nn
from torch.autograd import Variable
from torch.utils.data import DataLoader
import argparse
import time
import datetime
import os
import math
import tifffile as tiff
import numpy as np
import random
#############################################################################################################################################
parser = argparse.ArgumentParser()
parser.add_argument("--n_epochs", type=int, default=30, help="number of training epochs")
parser.add_argument('--GPU', type=str, default='0', help="the index of GPU you will use for computation (e.g. '0', '0,1', '0,1,2')")
parser.add_argument('--patch_x', type=int, default=128, help="patch size in x and y")
parser.add_argument('--patch_t', type=int, default=128, help="patch size in t")
parser.add_argument('--overlap_factor', type=float, default=0.5, help="the overlap factor between two adjacent patches")
parser.add_argument('--train_datasets_size', type=int, default=6000, help='How many patches will be used for training.')
parser.add_argument('--datasets_path', type=str, default='datasets', help="dataset root path")
parser.add_argument('--pth_path', type=str, default='./pth', help="the root path to save models")
parser.add_argument('--datasets_folder', type=str, default='./train', help="A folder containing files for training")
parser.add_argument('--output_path', type=str, default='./results', help="output directory")
parser.add_argument('--lr', type=float, default=0.0001, help='initial learning rate')
parser.add_argument("--b1", type=float, default=0.5, help="Adam: bata1")
parser.add_argument("--b2", type=float, default=0.999, help="Adam: bata2")
parser.add_argument('--select_img_num', type=int, default=10000000000, help='How many frames will be used for training.')
parser.add_argument('--test_datasize', type=int, default=10000000000, help='How many frames will be tested.')
parser.add_argument('--scale_factor', type=int, default=1, help='the factor for image intensity scaling')
opt = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = opt.GPU
#############################################################################################################################################
from SRDTrans import SRDTrans
from data_process import train_preprocess_lessMemoryMulStacks, trainset
from utils import save_yaml_train
from sampling import *
########################################################################################################################
# use isotropic patch size by default
opt.patch_y = opt.patch_x # the height of 3D patches (patch size in y)
opt.patch_t = opt.patch_t # the length of 3D patches (patch size in t)
opt.gap_x = int(opt.patch_x*(1-opt.overlap_factor)) # patch gap in x
opt.gap_y = int(opt.patch_y*(1-opt.overlap_factor)) # patch gap in y
opt.gap_t = int(opt.patch_t*(1-opt.overlap_factor)) # patch gap in t
opt.ngpu = opt.GPU.count(',')+1
opt.batch_size = opt.ngpu # By default, the batch size is equal to the number of GPU for minimal memory consumption
print('\033[1;31mTraining parameters -----> \033[0m')
print(opt)
########################################################################################################################
if not os.path.exists(opt.output_path):
os.mkdir(opt.output_path)
current_time = opt.datasets_folder+'_'+datetime.datetime.now().strftime("%Y%m%d%H%M")
output_path = os.path.join(opt.output_path, current_time)
pth_path = os.path.join('pth', current_time)
print("ckp is saved in {}".format(pth_path))
if not os.path.exists(pth_path):
os.mkdir(pth_path)
train_name_list, train_noise_img, train_coordinate_list, stack_index = train_preprocess_lessMemoryMulStacks(opt)
yaml_name = os.path.join(pth_path, 'para.yaml')
save_yaml_train(opt, yaml_name)
########################################################################################################################
L1_pixelwise = torch.nn.L1Loss()
L2_pixelwise = torch.nn.MSELoss()
denoise_generator = SRDTrans(
img_dim=opt.patch_x,
img_time=opt.patch_t,
in_channel=1,
embedding_dim=128,
num_heads=8,
hidden_dim=128*4,
window_size=7,
num_transBlock=1,
attn_dropout_rate=0.1,
f_maps=[8, 16, 32, 64],
input_dropout_rate=0
)
param_num = sum([param.nelement() for param in denoise_generator.parameters()])
print('\033[1;31mParameters of the model is {:.2f} M. \033[0m'.format(param_num/1e6))
if torch.cuda.is_available():
denoise_generator = denoise_generator.cuda()
denoise_generator = nn.DataParallel(denoise_generator, device_ids=range(opt.ngpu))
print('\033[1;31mUsing {} GPU(s) for training -----> \033[0m'.format(torch.cuda.device_count()))
L2_pixelwise.cuda()
L1_pixelwise.cuda()
########################################################################################################################
optimizer_G = torch.optim.Adam(denoise_generator.parameters(),
lr=opt.lr, betas=(opt.b1, opt.b2))
########################################################################################################################
cuda = True if torch.cuda.is_available() else False
Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
prev_time = time.time()
########################################################################################################################
time_start=time.time()
# start training
def train_epoch():
global prev_time
denoise_generator.train()
train_data = trainset(train_name_list, train_coordinate_list, train_noise_img, stack_index)
trainloader = DataLoader(train_data, batch_size=opt.batch_size, shuffle=True, num_workers=4)
for iteration, noisy in enumerate(trainloader):
noisy = noisy.cuda()
mask1, mask2, mask3 = generate_mask_pair(noisy)
noisy_sub1 = generate_subimages(noisy, mask1)
noisy_sub2 = generate_subimages(noisy, mask2)
noisy_sub3 = generate_subimages(noisy, mask3)
noisy_output = denoise_generator(noisy_sub1)
loss2neighbor_1 = 0.5*L1_pixelwise(noisy_output, noisy_sub2) + 0.5*L2_pixelwise(noisy_output, noisy_sub2)
loss2neighbor_2 = 0.5*L1_pixelwise(noisy_output, noisy_sub3) + 0.5*L2_pixelwise(noisy_output, noisy_sub3)
################################################################################################################
optimizer_G.zero_grad()
# Total loss
Total_loss = 0.5 * loss2neighbor_1 + 0.5 * loss2neighbor_2
Total_loss.backward()
optimizer_G.step()
################################################################################################################
batches_done = epoch * len(trainloader) + iteration
batches_left = opt.n_epochs * len(trainloader) - batches_done
time_left = datetime.timedelta(seconds=int(batches_left * (time.time() - prev_time)))
prev_time = time.time()
if iteration % 1 == 0:
time_end = time.time()
print(
'\r[Epoch %d/%d] [Batch %d/%d] [Total loss: %.2f] [ETA: %s] [Time cost: %.2d s] '
% (
epoch + 1,
opt.n_epochs,
iteration + 1,
len(trainloader),
Total_loss.item(),
time_left,
time_end - time_start
), end=' ')
if (iteration + 1) % len(trainloader) == 0:
print('\n', end=' ')
################################################################################################################
# save model
if (iteration + 1) % (len(trainloader)) == 0:
file_save_name = 'E_' + str(epoch + 1).zfill(2) + '_Iter_' + str(iteration + 1).zfill(4) + '.pth'
model_save_name = os.path.join(pth_path, file_save_name)
if isinstance(denoise_generator, nn.DataParallel):
torch.save(denoise_generator.module.state_dict(), model_save_name) # parallel
else:
torch.save(denoise_generator.state_dict(), model_save_name) # not parallel
for epoch in range(0, opt.n_epochs):
train_epoch()